Benchmarking Auto‐Regressive Spatio‐Temporal Predictions of Wildfire Spread From the FireBench Simulation Data Set With a Machine Learning Competition

Abstract As wildfire severity increases, predicting wildfire spread is important for fire and landscape management. However, wildfire spread is a complex phenomenon that is challenging to capture because it depends on many factors, such as wind speed, slope, landscape topography, and fuel properties. Machine Learning (ML) models are powerful tools that can process complex multi‐dimensional data, which makes them attractive for predicting wildfire spread. An ML competition was hosted to benchmark different ML architectures and techniques for the challenge of predicting dynamics of wildfire spread based on the FireBench wildfire data set. This data set considers different wind speeds and slopes, which are two of the main factors in wildfire propagation predictions. The challenge aims to benchmark robust ML models capable of predicting the fire line location over a duration of 20 time steps, corresponding to 20 s, and the models are further assessed after the competition for a total of 60 time steps to evaluate their extrapolation performance. Following the competition, generalizability of the models was tested on additional unseen conditions, involving different slopes and wind speeds. The ML architectures employed included convolutional neural networks, convolutional long short‐term memory, transformers, and UNets. It was found that many of these ML models not only improve fire spread predictions compared to a linear model and a baseline ML model which uses ConvLSTM and the UNet architecture, but also incorporate innovations such as thresholding or k‐fold cross‐validation to avoid autoregressive issues like prediction of spurious fire lines.